Two startups are looking to give enterprises and HPC laboratories something they are thirsty for in their compute-cluster environments: more bandwidth and less latency.
OctigaBay Systems Corp. last week unveiled the OctigaBay 12K, a Linux-based high-performance computing platform powered by Advanced Micro Devices Inc.s 64-bit Opteron processors that OctigaBay officials say will be 30 times faster than Gigabit Ethernet interconnect technology.
For its part, S2io Inc. has rolled out its Xframe server/ storage adapter, a 10 Gigabit Ethernet PCI-X device aimed at networked servers and storage devices. Officials at the Cupertino, Calif., company said the adapter, which will ship this quarter, will offer eight times the throughput of other interconnect technologies while cutting latency in half.
The Xframe is a key part of an evolving ecosystem surrounding 10 Gigabit Ethernet, company officials said. In servers, the adapter can be placed in a standard PCI-X slot and will enable administrators to replace six to eight Gigabit Ethernet ports with one 10 Gigabit Ethernet port. In addition, cabling for the systems will remain the same. For storage systems, the Xframe will improve network-attached storage speeds by up to five times over existing 2G-bps Fibre Channel technologies, they said.
OctigaBays challenge was finding a way to aggregate processing power within HPC environments while offering a high-speed interconnect and holistic systems management software. The obstacle to performance in HPC environments has been getting data to processors. Traditional environments use shared memory or I/O buses for clustering.
OctigaBay 12K includes up to 12 Opterons and an HPC-optimized Linux operating system. The systems RapidArray Interconnect technology includes 12 communications processors, and a 1-terabit-per-second switch fabric ties the Opterons to the fabric, offering up to 8G-bps bandwidth with 1 microsecond of latency. There also are six field programmable gate array coprocessors to increase application performance, as well as management hardware and software.
A single 12K shelf offers up to 58 gigaflops of computing power, said company officials, in Burnaby, British Columbia.
Eugene Fiume, chairman of the University of Torontos Department of Computer Science, said that the 12K stakes out the middle ground in HPC between the more expensive and limited symmetric multiprocessing systems and clusters of smaller servers, which have bandwidth and latency problems. The 12Ks ability to interoperate with other environments—thanks to its use of x86 processors, Linux and Message Passing Interface—will make it attractive to organizations running HPC systems.
“Most people put together high-performance systems statically,” said Fiume, who expects to buy a 12K. “They look at how theyre going to scale their systems in the future.”